Despite deep learning’s high precision in emotion identification, centralized training is associated with privacy and scalability concerns. The privacy-preserving federated learning model, Federated Hybrid-Optimized Emotion Recognition (Fed-HOER), introduced in this paper is an auto-tuning hyperparameters optimizer based on a hybrid Dung Beetle Optimizer-Fick’s Law Algorithm (DBO-FLA) optimizer. The global and local searches are optimized at two levels, and validation loss is minimized by 22%–24% without sharing raw data. The experiments on Extended Cohn–Kanade (CK+), Japanese Female Facial Expressions (JAFFE), and Karolinska Directed Emotional Faces (KDEF) exhibit a high generalization rate with a mean accuracy of 98.14. The findings demonstrate that Fed-HOER is statistically significantly better than baseline configurations. The results show that the suggested framework offers a favorable trade-off between predictive accuracy and privacy protection, which is why it can be used in the healthcare, educational, and other emotion-related fields.